Some global properties of neural networks. L. Accardi and A. Aiello. Laboratorio di Cibernetica del C.N.R., Arco Felice (Na), Italy

Transcription

1 Some global properties of neural networks L. Accardi and A. Aiello Laboratorio di Cibernetica del C.N.R., Arco Felice (Na), Italy 1

2 Contents 1 Introduction 3 2 The individual problem of synthesis 4 3 The Global Problem of Synthesis 7 2

3 Abstract. Some properties of the global behaviour of a model of neural network are considered. The geometric concept of quadrant degeneration is studied and it is shown to be independent of the algebraic concept of rank degeneration. The results obtained are employed to solve some global problems of synthesis (i.e., independent of the initial state of the network) without the use of the theory of linear inequalities. 1 Introduction The mathematical model of a neural network with which we will be concerned has been introduced by Caianiello (1961). In this model the evolution of a network is described by means of Neuronic Equations (N.E.) which describe the instantaneous activity of the net, and Mnemonic Equations (M.E.) describing the learning processes. The Adiabatic Learning Hypothesis which asserts that learning processes take place more slowly than the processes described by the neuronic equations, allows to study independently the N.E. and the M.E. within a time interval, whose width is discussed by Caianiello (1961). We will here be concerned exclusively with the study of neuronic equations in the special form obtained under self duality conditions which characterize the so called normal systems (Caianiello, 1966). In Section 1, after a brief recall of the formalism, we consider the individual problem of synthesis and we show the equivalent, for this case, of a well known theorem enables us to give some helpful criterion to solve concrete problems. In Section 2 the global problem of synthesis is stated and a result is proved which generalizes a theorem given in Accardi (1971). Moreover, there are some practical considerations of the synthesis of reverberations whose lengths must not cross preassigned upper or lower bounds. Finally, in Section 3 the solution of some concrete problems of synthesis are presented. An example which shows the independence between rank degeneration and quadrant degeneration is also given. 3

4 2 The individual problem of synthesis We consider the formal neurons as binary elements which can assume the values 1 or 1. The states of the network are vectors, whose i th components represent the state of the i th neuron. Time is quantized in the mode. In the case of normal systems the state σ t+1 of an N neuron network at the time t + 1 can be expressed by the (vector) equation: σ t+1 = sgn (Aσ t ) (1) where A is an N N matrix, and sgn [ ] is the signum function, defined componentwise on R {0} by: sgn x = { +1 if x > 0 1 if x < 0 (2) The problem of synthesis of neural networks may be expressed in the following manner. Given a sequence σ,..., σ R+1 of states, determine a matrix A such that σ i+1 = sgn (Aσ i ), 1 i R (3) We will refer to this as the individual problem of synthesis, that is, the network is required to realize a single preassigned sequence of states. If σ R+1 = σ 1 the sequence will be called a reverberation ; if σ l σ m for l < m we will say that the sequence belongs to a transient. It may be useful to state the equivalent for this model of a well known result about threshold functions and some of its consequences (Elgot, 1960). Denoting with, the scalar product which maps the vectors x = (x i ); y = (y i ) into the number x, y = N x i y i and denoting with a i the i th row i=1 of the matrix A, then Eq. (3), written componentwise, becomes (σ ij being the j th component of the vector σ i ): one can thus define the sets σ i+1,j = sgn ϱ j, σ i ; 1 i R ; 1 j N (4) I + j = {σ i i R; σ i+1,j = +1} I j = {σ i i R; σ i+1,j = 1} 4

5 (5) we then have I + j I j = and from I + j I j = {σ 1,..., σ R } for every j N, Eqs. (3) can be written: if σ i I j N σ i, a j > 0 if σ i I + j σ i, a j < 0 ; i j (6) Defining the sets, for 1 j N I j = {σ σ I j } ; I j = I + j ( I j ) (7) Eqs. (6) can be written in compact form: σ i, a j > 0 ; σ i I j ; 1 j N (8) As a consequence, a matrix A satisfying our problem exists if and only if the N systems of linear inequalities σ i, x > 0 ; σ i I j ; 1 j N (9) are consistent. A well known theorem of convex analysis (Rockafellar, 1969) asserts that this is the case if and only if the null vector does not belong to the convex hull of the set I j for 1 j N, that is, if there exist no R non negative numbers t σ whose sum equals 1, such that, for some j, t σ σ = 0 (10) σ I j This is equivalent to t σ u σ = (11) σ I j where u σ = 1 2 (σ + 1), i.e., u σ is the boolean state associated to σ. Introducing the matrix u j = [u 1,..., u R ], σ i = 2u i 1 I j, Eq. (11) is equivalent to: u j t = (12) Eqs. (11), (12) provide, through classical theorems of linear algebra, some simple sufficient criteria of realizability. For example, if (u i ) 1 i K is a generating system for the set 2I j 1, i.e., a maximum number of independent 5

6 vectors, a sufficient condition for realizability is that the vector 1 does not belong to the space generated by the u i. Suppose now that the sequence (σ) 1 i R+1 is realizable, i.e., there exists a network which contains this sequence in its evolution. If σ is an arbitrary vector of Q N, then, from Eq. (10) it follows that the sequence (σ i ) 1 i R+2 ; σ R+2 = σ, will be realizable if and only if no vectors λ, with negative components exist, such that for some j = 1, 2,..., N, the equation S j λ = τ (13) where the R columns of the matrix S j are the vectors of I j in the preceding notations, and τ = ±σ R+1 according to the sign of the j th component of σ R+2. The geometric meaning of condition (13) can be expressed in this way: denote P j σ the j th component of the state σ Q N ; then the sequence (σ i ) 1 i R+2 can be realized if and only if for every j = 1,..., N the vector (P j σ)σ R+1 = τ j does not belong to the convex cone generated by the set I j (see (7)); this means that for no positive numbers (λ σ ) σ I j one has: λ σ σ = τ j (14) σ I j taking scalar products for τ j on both sides, on finds λ σ σ, τ j = N (15) σ I j so that a sufficient condition for the realizability of the sequence (σ i ) 1 i R+2 is that for every j = 1, 2,..., one has: σ, τ j 0 ; σ I j (16) Suppose now that the sequence (σ i ) 1 i R is N net realizable. Then defining in the above notations the sets Γ(I j); 1 j N, as the closed convex hull of I j, the following proposition answers the question. Proposition 1 If the sequence (σ i ) 1 i R is N net realizable, and ν is the number of indices j = 1,..., N, such that the line y = λ R, does not intersect the set Γ(I j), then there exist 2 ν vectors σ Q N such that the sequence {(σ i ), σ} is N net realizabe. 6

7 In fact, from Eq. (10) one can see that the sequence {(σ i ), σ} is N net realizable if and only if for every j = 1,..., N vector λ(p j σ)σ R does not belong to Γ(I j) where λ < 0 and P j σ is the j th component of σ. So, if the straight line y = λσ R does not intersect Γ(I j), then the j th component of σ ca be arbitrary. If λσ R belongs to Γ(I j) for some λ, then λσ R cannot belong to T (I j) since (σ i ) 1 i R is realizable. Thus our hypothesis is equivalent to saying that ν components of σ are arbitrary. 3 The Global Problem of Synthesis The formulation of the problem of synthesis given in the preceding paragraph depends on the initial state: i.e., if a matrix A is found which realizes a preassigned sequence of states (σ i ) 1 i R+1, nothing in general is known about the evolution of the network when an initial state τ is chosen different from any of the σ i in the sequence. There are, nevertheless, some problems in which one is mainly interested in the final stable configuration of the network resulting from an arbitrary initial excitation. One can refer, for example, to Caianiello et al. (1967) for a theoretical statement of the problem, and to Burattini and Liesis (1970) for an analysis of this kind of situation from an experimental point of view. This problem can clearly be reduced to the preceding one by the consideration of a sufficiently large number of systems of linear equalities, but this technique soon becomes very involved, even for small N (number of neurons) and R (length of reverberations). To handle this problem from an intrinsic standpoint, we want to develop a geometric formulation of it. We denote Q N the set of states of the network, identified, as usual, with the set of vertices of the unit symmetric cube in R N, and define the mapping T A : σ Q N sgn (Aσ) Q N (17) the evolution equation of the network is then written T A (σ i ) = σ i+1 (18) and a generic state σ belongs to a reverberation of length R if and only R is the smallest positive integer such that: T R A (σ) = σ ; where T 0 A(σ) = σ; T n+1 A = T A (T n A) (19) 7

8 If, on the contrary, for every natural integer R one has TA R (σ)not = σ the state σ is said to belong to a transient. The function T A has the following two properties: (i) T A is symmetric, i.e., T A ( σ) = T A (σ) for every σ in Q N. (ii) T A maps Q into itself: i.e., T A Q N Q N. Since Q N is a finite set, and, for every integer n one has: T n+1 A (QN ) T n A(Q N ) (20) the inclusion can be proper only for a finite number of steps. Let m be the least integer such that equality holds in (20) and set T m A (Q N ) = V (21) the subset QV of Q N thus defined, is the largest among the subsets of Q N on which the function T A acts as a permutation. This will be called the core of the network whose matrix is A and is characterized by the following property: a state belongs to a reverberation of the network if and only if it belongs to V. In analogy with Q N, the points of V may be considered as the vertices of a polyhedron (in general not a cube!). Like Q N, the polyhedron V is symmetric because such is the map T A. We can sum up the above consideration in this way: to every self dual network A corresponds a symmetric subset V of Q N such that T A acts as a permutation on the elements of V, and V is the largest subset of Q with this property. We can express the global problem of synthesis as follows: given a symmetric subset V of Q N and a permutation P on the points of V, does a self dual network A exist, such that V is the core of A and the permutation induced by T A on V coincides with P? The core of a network is its final stable configuration since starting from an arbitrary state after a finite number of steps ( m) the network decays into a state belonging to the core which is stable with respect to the evolution of the net, in the sense that if the initial state belongs to the core, all the subsequent states of the network also belong to it. Thus the problem of vector realizability of networks can be generalized as follows: given a set S and an action on S of a subgroup G of the group of permutations on elements of S, we look for the minimum integer N such that there exists a one to one map ϕ of S into a polyhedron P of K N (K 8

9 an arbitrary field) and an homomorphism ϱ of G into the group of transformations (not necessarily linear) of P into itself such that if g G one has ϕ g = ϱ(g) ϖ. This approach is by no means restricted to the case of autonomous threshold nets, but extend to the general problem of vector realizations of automata. This formulation allows a mathematical approach to the problem of realizability of nets (i.e., systems of boolean functions) which is different from the usual ones. As long as we are concerned with the reverberation behavior of the net we see that the central concept involved is that of action of the permutation group on the vertices of a polyhedron. This action is usually expressed by means of the Heaviside function composed with an affine mapping, denoting with ϕ the transformation which maps the boolean cube onto the unit symmetric one, with H the action of the Heaviside function, with S the action of the Heaviside function, with S the action of the signum function one obtains the equality: ϕh = Sϕ (22) The transformation ϕ has two peculiar geometric features: 1) it preserves dimension; 2) it maps a cube onto a cube. The meaning of these properties is simple: the transition from one model to the other preserves both the number of neurons and the code. It is apparent that in equality (22) every trace of these properties is lost, so that this equation can be assumed (ϕ being one to one) as the definition of equivalence of two vector models of the same neural net. More precisely: if V is an arbitrary polyhedron of a space K N and S is an action of a subgroup of a permutation group, we say that the couple (V, S) is an equivalent vector model of a couple (Q, H) if there exists a map ϕ : Q N V such that ϕh = Sϕ. As an example of global property of a network we consider the following proposition which extends a recent result of one of the authors. Proposition 2 Let A be an N network of rank K, then, if we denote with (R i ) 1 i m the lengths of all its reverberations, we have m R i 2 N 2 N K (23) i=1 Proof. Since A has rank K, it projects all the points of Q N (states of the network) on a K dimensional linear manifold. It ha been shown (Accardi, 9

10 1971) that such a manifold can intersect at most 2 N 2 N K+1 +2 quadrants so that this will be a fortiori an upper bound for the number of points belonging to the core (which must belong to one of these quadrants ). This number, because of the characteristic property of the core stated previously, is more than the sum of the lengths of all the reverberations of the network. The limitation expressed in (??) is the best possible one, depending only on the rank of the matrix. We refer, for this and other considerations, to the above mentioned paper in which an example is given where equality holds in (??). Thus the rank of a network provides an upper bound for the number of states in its core. It can be shown (see Section 3) that this is not the case for a lower bound of this number; therefore it may be of interest to know some type of sufficient conditions which supply such a lower bound. This will ensure, as in the following example, that the network is sufficiently rich in reverberations, either in number or in length. Let, for simplicity, N = 2 r. In such a case one can choose from Q an orthogonal basis (σ i ) 1 i N (this is trivially verified by induction on the tensor product of q copies of R 2 ). Put K = 2 q with p < q and S = [σ 1,..., σ N ]; S 1 = [ σ 1,..., σ K ; s K+11,..., s N ] where s j is an arbitrary linear combination of the (σ i ) 1 i K for j = K + 1,..., N. Now define A by AS = S 1 ; A = 1 N S is T (24) Then clearly A has the rank K; its core contains the ν vertices of Q N which are linear combinations of the (σ i ) 1 i K and the network A acts on each of these vertices as a non linear oscillator. Obviously a similar argument may hold for any matrix of the type AB where B is a matrix which leaves invariant the polyhedron Q(K) generated by the (σ i ) 1 i K. For any such matrix, denoted by (R i ) the set of the lengths of all the reverberations, one has: ν R i 2 N 2 N K (25) i ( ) N (for example if K = N 1 then ν = ). N/2 The upper bound given in Proposition 2 for the number of states of the core of a network of rank K can be improved with some additional hypothesis on the matrix A of the network. We consider here one of these cases as 10

11 an illustration of the method. Suppose x T U = 0 is the equation of the K dimensional subspace defined by A. The matrix U (which is of order N (N K) and rank N K can be decomposed into: U = [ ] B V B where B is of rank N K and the above equation is equivalent to the following: x T = [ y T V y T ] (26) where y is an arbitrary vector with K components. We want to look for the number of signs obtainable with vectors x of the form (??) when the matrix V is such that in each of its columns there may occur some zero and, if v (j) denotes its j th column, for every couple of indices h and i there exist at least two components, say the α th and the β th, non null both in v (h) and v (i) and such that: sgn v α (i) = sgn v α (h) sgn v (i) β = sgn v (h) β (27) (this is surely possible if, for instance, 2 K 1 N K). Then, if ν h is the number of zeros in the column v (h), the set M h of all the σ Q K such that the sign of the product x, v (h) does not depend on the x i s when ever sgn x = σ, contains exactly 2 ν h+1 elements. Furthermore M h M i = if h i, in fact, because of our hypothesis there exist two components α and β such that if then, necessarily: sgn (x α ) sgn (v α (h) ) = sgn (x β ) sgn (v (h) β ) (28) sgn (x α ) sgn (v (i) α ) sgn (x β ) sgn (v (i) β ) (29) which means that if sgn x = σ M h, then the product x, v (i) depends on the x i s. Now for every σ Q K, let h σ be the number of columns v (j) in V such that the product x, v (j) does not depend on the x i s whenever sgn x = σ. Our hypothesis implies that h σ = 1 or h σ = 0. Then the vector x T V can give rise to at most 2 N K hσ different signs, so that, when σ ranges in Q K the number 2 N K hσ (30) σ Q K 11

12 is an upper bound for the number of vertices in the core of the network. We have, then, N K 2 hσ = 2 hσ + 2 hτ (31) σ Q K j=1 σ M j τ L where L denotes the set of the σ Q K not belonging to any of the M j, i.e., such that if sgn x = σ the product x, v (i) depends on the x j s for every i = 1, 2,..., N K. Since the M j are disjoint and each of them contains 2 νj+1 elements, then 2 hτ = 2 K τ L but if σ M j then h σ = 1 and therefore which yields N K j=1 σ Q K 2 N K hσ σ M j 2 hσ= N K j=1 N K 1 2 j=1 = 2 N 3 2 2N K 2 ν j+1 (32) 2 ν j (33) ( N K j=1 2 ν j ) (34) It has already been observed (see Accardi, 1971) that besides the rank degeneration which has been investigated in Proposition 2 there may arise another kind of degeneration, called quadrant degeneration which takes place when different vertices of Q N are mapped into a single quadrant of R N so that the signum function, when applied to them, maps them all into the same vertex of Q N. The following example shows that these two kinds of degeneration are independent within the limits of Proposition 2 and that the essential feature of the equivalence of nets is topological stability of the evolution within some tolerance cone - rather than algebraic - rank invariance. Moreover, this example solves, without the introduction of any controlling element, the problem, examined in Caianiello et al. (1967), of synthesizing a net which has the same immediate reaction regardless of the impinging excitation. We are thus looking for a matrix A such that for a preassigned τ Q N sgn Aσ = ±τ ; σ Q N (35) 12

13 i.e., A must map the whole symmetric cube into the quadrants defined by τ and τ. And a sufficient condition for this is that A maps Q N into the cone C(τ) of the axis the direction of τ and tangent to the walls delimiting the quadrant identified by τ itself. Denoting θ j (τ) the angle between τ and its projection on the j th wall of the τ quadrant, then for j = 1,..., N, cos θ j (τ) = τ, τ ( τ, e j ) 2 N(N 1) = N 1 = 1 1 N(N 1) N Therefore a vector x belongs to the cone C(τ) if cos( xτ) (36) 1 1 N. Thus, choosing A such that τ is an eigenvector of A T corresponding to the positive eigenvalue λn one has, Aσ, τ = σ, A T τ = λn σ, τ (37) and if N is odd cos(a σ τ) = Aσ, τ Aσ τ = λn σ, τ Aσ τ λn A N = λ A where A is a norm for the matrix A satisfying the inequality Ax A x see, for example, Gantmacher (1964) and the required conditions on A are (N being odd): A T τ = λn τ λ\ A 1 1 N It is instructive to give a geometrical interpretation of this problem. Suppose first, as is always possible, that the matrix A in (??) is positive definite. Then it is well known (see Gantmacher, 1964) that its action on the vectors of R N consists of a stretching of their components along N orthogonal directions (the eigen axis of the matrix). Condition (??) can therefore be interpreted as imposing that the stretching along the axes of τ must be much bigger than along other axes of A. In particular, we can choose A in (??) as any positive (not necessarily definite) matrix of arbitrary rank K (1 K N) which shows that the essential feature for the evolution of a network A is the geometrical action of A on the points of Q and not the rank of A. 13

14 Acknowledgments. The authors are grateful to Professor E.R. Caianiello for useful discussions. References [1] Accardi L.: Rank and reverberations in neural networks, Kybernetik 8, (1971). [2] Arimoto T.: Periodical sequences realizable by an autonomous net of threshold elements, papers of the Committee on Automaton and Automatic Control, Inst. Elec. Commun. of Japan, March [3] Burattini E., Liesis V.: To appear. [4] Caianiello E.R.: Outline of a theory of thought processes and thinking machines, J. theor. Biol. 2, (1961). [5] Burattini E., Liesis V.: Decision equations and reverberations, Kybernetik 3, (1966a). [6] Burattini E., Liesis V.: A study of neural networks and reverberations, Bionics Symposium, Dayton, Ohio (1966b). [7] Luca A. de, Ricciardi L.M.: Reverberations and control of neural networks, Kybernetik 4, (1967). [8] Elgot: Switching circuit theory and logical design, p [9] Gantmacher: Matrix Theory, Chelsea [10] Rockafellar: Convex analysis, Princeton